With the rapid advancement of accelerator technologies, pre-training large language models (LLMs) on heterogeneous accelerator clusters has become increasingly crucial for maximizing hardware utilization. Existing systems, however, suffer from inaccurate training time modeling, which undermines the parallelization optimizations built upon it. Moreover, for current approaches, the vast configuration search space makes exhaustive exploration infeasible, forcing a trade-off between search time and training efficiency. To overcome these limitations, we introduce Poseidon, an efficient and scalable LLM training framework designed with heterogeneity awareness. Its core is an explicit training time model based on a directed acyclic graph. Building on this graph, Poseidon employs two efficient, theoretically grounded strategies: stage-level pruning via early stopping with partial estimation, and layer-to-stage mapping exploiting a ridge-like distribution pattern. These strategies reduce the search space without sacrificing optimal training efficiency. Experiments on heterogeneous clusters show that Poseidon improves training throughput by up to $2.76\times$ over state-of-the-art systems.
展开完整摘要收起摘要↓
With the rapid advancement of accelerator technologies, pre-training large language models (LLMs) on heterogeneous accelerator clusters has become increasingly crucial for maximizing hardware utilization. Existing systems, however, suffer from inaccurate training time modeling, which undermines the parallelization optimizations built upon it. Moreover, for current approaches, the vast configuration search space makes exhaustive exploration infeasible, forcing a trade-off between search time and training efficiency. To overcome these limitations, we introduce Poseidon, an efficient and scalable LLM training framework designed with heterogeneity awareness. Its core is an explicit training time model based on a directed acyclic graph. Building on this graph, Poseidon employs two efficient, theoretically grounded strategies: stage-level pruning via early stopping with partial estimation, and layer-to-stage mapping exploiting a ridge-like distribution pattern. These strategies reduce the search space without sacrificing optimal training efficiency. Experiments on heterogeneous clusters show that Poseidon improves training throughput by up to $2.76\times$ over state-of-the-art systems.
Few-shot fine-grained object detection (FGOD) in remote sensing imagery is challenging because limited annotations must support both object localization and discrimination among visually similar subcategories. Although multimodal large language models (MLLMs) provide strong coarse object localization, they lack explicit visual evidence for reliable fine-grained recognition. To address this limitation, we propose ProtoRAG, a prototype-based retrieval-augmented framework that decouples coarse localization from fine-grained recognition by equipping MLLMs with an external object-level visual memory. To construct a reliable visual memory from limited support samples, we introduce Discriminative Prototype Space Learning (DPSL), which encourages discriminative and prototype-stable representations through supervised contrastive learning and prototype-consistency regularization. We further develop an uncertainty-guided candidate-constrained reasoning strategy that augments MLLMs with retrieved candidate-specific visual references and invokes multimodal reasoning only for ambiguous instances. Extensive experiments show that ProtoRAG consistently surpasses representative baselines in nine few-shot settings, outperforming the strongest baselines by 14.80, 2.27, and 4.04 mAP$_{50}$ on MAR20, HRSC2016, and FAIR1M-2.0, respectively.
展开完整摘要收起摘要↓
Few-shot fine-grained object detection (FGOD) in remote sensing imagery is challenging because limited annotations must support both object localization and discrimination among visually similar subcategories. Although multimodal large language models (MLLMs) provide strong coarse object localization, they lack explicit visual evidence for reliable fine-grained recognition. To address this limitation, we propose ProtoRAG, a prototype-based retrieval-augmented framework that decouples coarse localization from fine-grained recognition by equipping MLLMs with an external object-level visual memory. To construct a reliable visual memory from limited support samples, we introduce Discriminative Prototype Space Learning (DPSL), which encourages discriminative and prototype-stable representations through supervised contrastive learning and prototype-consistency regularization. We further develop an uncertainty-guided candidate-constrained reasoning strategy that augments MLLMs with retrieved candidate-specific visual references and invokes multimodal reasoning only for ambiguous instances. Extensive experiments show that ProtoRAG consistently surpasses representative baselines in nine few-shot settings, outperforming the strongest baselines by 14.80, 2.27, and 4.04 mAP$_{50}$ on MAR20, HRSC2016, and FAIR1M-2.0, respectively.
Low-rank adaptation (LoRA) has become the standard for parameter-efficient fine-tuning of large language models. Most LoRA variants follow a uniform-LR convention, applying a single global learning rate across every rank-one component of every adapter. We show that this convention overlooks substantial within-module heterogeneity, where the rank-one components of a LoRA adapter update at highly uneven rates and low-velocity modules converge to concentrated singular spectra that underutilize the nominal rank budget. To address this, we propose an adaptive anisotropic learning-rate model that assigns each rank-one component its own effective learning rate, computed online from training-time signals and mean-normalized per module to preserve the global LR budget. AnLR-LoRA instantiates this model with two signals available during AdamW optimization, namely function-space velocity and Adam SNR, as a lightweight scheme with no extra trainable parameters. Across commonsense reasoning, natural language generation and visual instruction-tuning benchmarks, AnLR-LoRA consistently improves over LoRA while encouraging broader use of rank capacity, with gains that remain robust across a wide range of global learning rates and transfer cleanly to other LoRA variants.
展开完整摘要收起摘要↓
Low-rank adaptation (LoRA) has become the standard for parameter-efficient fine-tuning of large language models. Most LoRA variants follow a uniform-LR convention, applying a single global learning rate across every rank-one component of every adapter. We show that this convention overlooks substantial within-module heterogeneity, where the rank-one components of a LoRA adapter update at highly uneven rates and low-velocity modules converge to concentrated singular spectra that underutilize the nominal rank budget. To address this, we propose an adaptive anisotropic learning-rate model that assigns each rank-one component its own effective learning rate, computed online from training-time signals and mean-normalized per module to preserve the global LR budget. AnLR-LoRA instantiates this model with two signals available during AdamW optimization, namely function-space velocity and Adam SNR, as a lightweight scheme with no extra trainable parameters. Across commonsense reasoning, natural language generation and visual instruction-tuning benchmarks, AnLR-LoRA consistently improves over LoRA while encouraging broader use of rank capacity, with gains that remain robust across a wide range of global learning rates and transfer cleanly to other LoRA variants.
Large language models now compute correct tax liabilities on over 90% of well-formed cases in statutory benchmarks, which makes them candidates for the tax-advisory and compliance systems that consume such an answer directly. Real legal inputs, however, are frequently defective: required facts are missing, or stated facts contradict one another. Accuracy on clean benchmarks says nothing about how a model behaves then, and a system that computes straight through a defective input returns a confident number with no sign that anything is wrong. This raises two questions: does a model asked to solve a case abstain when the input is defective, and when it does not, can the same model catch the defect when asked instead to verify the input? We study six recent models on SARA-derived tax cases under missing-fact and contradictory-fact perturbations. The strongest models abstain when a fact is missing but compute through injected contradictions, returning the clean-input answer 63-76% of the time with no signal of the conflict; asked instead to verify the same input, they flag most of those contradictions. We wire that verification call into a simple contradiction gate: one extra call that abstains when the model reports a conflict. Across all six models it recovers most of the missed contradiction abstention at a clean-accuracy cost of at most about 5 percentage points, with no training and no external tooling. High accuracy on well-formed inputs is therefore an incomplete measure of reliability, and the detection the solver misses is cheaply recoverable with a single self-check.
展开完整摘要收起摘要↓
Large language models now compute correct tax liabilities on over 90% of well-formed cases in statutory benchmarks, which makes them candidates for the tax-advisory and compliance systems that consume such an answer directly. Real legal inputs, however, are frequently defective: required facts are missing, or stated facts contradict one another. Accuracy on clean benchmarks says nothing about how a model behaves then, and a system that computes straight through a defective input returns a confident number with no sign that anything is wrong. This raises two questions: does a model asked to solve a case abstain when the input is defective, and when it does not, can the same model catch the defect when asked instead to verify the input? We study six recent models on SARA-derived tax cases under missing-fact and contradictory-fact perturbations. The strongest models abstain when a fact is missing but compute through injected contradictions, returning the clean-input answer 63-76% of the time with no signal of the conflict; asked instead to verify the same input, they flag most of those contradictions. We wire that verification call into a simple contradiction gate: one extra call that abstains when the model reports a conflict. Across all six models it recovers most of the missed contradiction abstention at a clean-accuracy cost of at most about 5 percentage points, with no training and no external tooling. High accuracy on well-formed inputs is therefore an incomplete measure of reliability, and the detection the solver misses is cheaply recoverable with a single self-check.
The exponentiation of Artificial intelligence (AI) in the recent past has entered a transformative era that has been driven by the growth in large language models (LLMs), large-scale compute infrastructures, and autonomous reasoning systems. However, the rapid acceleration of AI has increasingly shown technological, societal, economic, ethical and infrastructural challenges associated with peak data limitations, rising computational demands, synthetic data recursion, valuation inflation, and societal instability. The traditional scaling paradigms that have powered the modern AI systems are gradually encountering friction in sustaining continuous exponential growth. This paper views ``the end of AI exponentiation,'' thus exploring how it flutters inside and outside the bubble, where instability emerges within the AI ecosystem through compute and data-center races, speculative investments, and the rat-race toward superintelligence, and outside the ecosystem through labor disruption, governance concerns, public uncertainty, and geopolitical acceleration surrounding future intelligent systems and infrastructures globally.
展开完整摘要收起摘要↓
The exponentiation of Artificial intelligence (AI) in the recent past has entered a transformative era that has been driven by the growth in large language models (LLMs), large-scale compute infrastructures, and autonomous reasoning systems. However, the rapid acceleration of AI has increasingly shown technological, societal, economic, ethical and infrastructural challenges associated with peak data limitations, rising computational demands, synthetic data recursion, valuation inflation, and societal instability. The traditional scaling paradigms that have powered the modern AI systems are gradually encountering friction in sustaining continuous exponential growth. This paper views ``the end of AI exponentiation,'' thus exploring how it flutters inside and outside the bubble, where instability emerges within the AI ecosystem through compute and data-center races, speculative investments, and the rat-race toward superintelligence, and outside the ecosystem through labor disruption, governance concerns, public uncertainty, and geopolitical acceleration surrounding future intelligent systems and infrastructures globally.
Multimodal sentiment analysis and emotion recognition in conversations demand effective modeling of heterogeneous interactions across textual, acoustic, and visual modalities. Although large language models (LLMs) offer powerful language understanding, adapting them to multimodal affective computing remains challenging: full-model fine-tuning is computationally prohibitive, while many existing lightweight adapters fail to preserve rich textual cues during cross-modal fusion. To address these limitations, we propose the multi-view text-guided multimodal fusion adapter (MVFA), a parameter-efficient framework that augments frozen LLMs with strong multimodal reasoning capability. MVFA first constructs complementary text views via max pooling, mean pooling, and attention pooling; these views then guide cross-modal interactions with audio and visual features. The fused multimodal representations are subsequently compressed into a compact set of learnable pseudo-tokens through an Enhanced Q-Former Fusion Module. Using ChatGLM3-6B-base as the primary backbone, we further validate MVFA on LLaMA2-7B and Qwen3-8B to examine its portability across multiple frozen LLM backbones. MVFA is evaluated on three challenging datasets: CH-SIMS V2.0, MELD, and CHERMA. Experimental results demonstrate that MVFA achieves state-of-the-art performance on key metrics while updating only a small fraction of parameters. Specifically, it attains 84.62% Acc2 and 84.59% F1 on CH-SIMS V2.0, 67.36% Acc and 66.03% WF1 on MELD, and 74.66% Acc on CHERMA. These findings establish multi-view text-guided fusion as an effective and scalable paradigm for parameter-efficient multimodal LLM adaptation in affective computing. The code is publicly available at https://github.com/Overwhelm1208/MVFA.
展开完整摘要收起摘要↓
Multimodal sentiment analysis and emotion recognition in conversations demand effective modeling of heterogeneous interactions across textual, acoustic, and visual modalities. Although large language models (LLMs) offer powerful language understanding, adapting them to multimodal affective computing remains challenging: full-model fine-tuning is computationally prohibitive, while many existing lightweight adapters fail to preserve rich textual cues during cross-modal fusion. To address these limitations, we propose the multi-view text-guided multimodal fusion adapter (MVFA), a parameter-efficient framework that augments frozen LLMs with strong multimodal reasoning capability. MVFA first constructs complementary text views via max pooling, mean pooling, and attention pooling; these views then guide cross-modal interactions with audio and visual features. The fused multimodal representations are subsequently compressed into a compact set of learnable pseudo-tokens through an Enhanced Q-Former Fusion Module. Using ChatGLM3-6B-base as the primary backbone, we further validate MVFA on LLaMA2-7B and Qwen3-8B to examine its portability across multiple frozen LLM backbones. MVFA is evaluated on three challenging datasets: CH-SIMS V2.0, MELD, and CHERMA. Experimental results demonstrate that MVFA achieves state-of-the-art performance on key metrics while updating only a small fraction of parameters. Specifically, it attains 84.62% Acc2 and 84.59% F1 on CH-SIMS V2.0, 67.36% Acc and 66.03% WF1 on MELD, and 74.66% Acc on CHERMA. These findings establish multi-view text-guided fusion as an effective and scalable paradigm for parameter-efficient multimodal LLM adaptation in affective computing. The code is publicly available at https://github.com/Overwhelm1208/MVFA.
作者Zeyu Wu, Junchao Wu, Shudong Liu, Runzhe Zhan, Xin Chen, Shu Yang, Yichao Du, Longyue Wang, Weihua Luo, Jinsong Su, Derek F. Wong
Existing Supervised Fine-Tuning paradigms, particularly Full Parameter Fine-Tuning are often plagued by parameter redundancy, inconsistent data quality, and catastrophic forgetting, which current methods typically address in isolation and lack a unified optimization signal to bridge data selection, parameter updates, and knowledge preservation. To address this, we propose Neuron-Guided Fine-Tuning (NGFT), a holistic framework that leverages neuron activation patterns as a universal proxy to unify the fine-tuning lifecycle. NGFT operates via three synergistic mechanisms: (1) Adaptive Task-Specific Neuron Selection, which identifies essential neurons in a single forward pass to concentrate updates and reduce redundancy; (2) Activation-Based Data Selection, which prioritizes information-dense samples that maximize contribution to key neurons; and (3) Neuron Activation Alignment, a novel loss function that anchors activations to pre-trained states, deepening representation learning and preserving general knowledge. Experimental results across three models across both domain-specific and general benchmarks demonstrate that NGFT significantly outperforms existing mainstream fine-tuning methods in both efficiency and performance, while effectively mitigating catastrophic forgetting.
展开完整摘要收起摘要↓
Existing Supervised Fine-Tuning paradigms, particularly Full Parameter Fine-Tuning are often plagued by parameter redundancy, inconsistent data quality, and catastrophic forgetting, which current methods typically address in isolation and lack a unified optimization signal to bridge data selection, parameter updates, and knowledge preservation. To address this, we propose Neuron-Guided Fine-Tuning (NGFT), a holistic framework that leverages neuron activation patterns as a universal proxy to unify the fine-tuning lifecycle. NGFT operates via three synergistic mechanisms: (1) Adaptive Task-Specific Neuron Selection, which identifies essential neurons in a single forward pass to concentrate updates and reduce redundancy; (2) Activation-Based Data Selection, which prioritizes information-dense samples that maximize contribution to key neurons; and (3) Neuron Activation Alignment, a novel loss function that anchors activations to pre-trained states, deepening representation learning and preserving general knowledge. Experimental results across three models across both domain-specific and general benchmarks demonstrate that NGFT significantly outperforms existing mainstream fine-tuning methods in both efficiency and performance, while effectively mitigating catastrophic forgetting.
Speculative Jacobi Decoding (SJD) is an important approach for accelerating autoregressive image generation. Although SJD has shown superior performance, recent studies point out that it usually suffers from a token ambiguity issue during token verification but its reason can not be well explained. To figure out this reason, in this paper, we conduct a visualization analysis on vision token and find that different from text tokens, vision tokens generally corresponds to some local, small, and unclear vision details, which means only using single token is difficult to accurately express a certain semantic, thereby causing token ambiguity issue. To this end, we propose a novel Speculative Jacobi Decoding with Semantics Verification (called SJD-SV), for accelerating autoregressive image generation. The key idea is that leveraging the strong correction characters between tokens to recognize semantic-aware token subsequence and then instead of perform token-by-token verification, turning to perform verification on semantic-aware token subsequence level for accelerating image generation. In particular, our method is plug-in, which can be directly integrated into existing SJD and its variants. Extensive experiments on various datasets show that existing SJD methods achieve significant performance improvement after integrating our SJD-SV method.
展开完整摘要收起摘要↓
Speculative Jacobi Decoding (SJD) is an important approach for accelerating autoregressive image generation. Although SJD has shown superior performance, recent studies point out that it usually suffers from a token ambiguity issue during token verification but its reason can not be well explained. To figure out this reason, in this paper, we conduct a visualization analysis on vision token and find that different from text tokens, vision tokens generally corresponds to some local, small, and unclear vision details, which means only using single token is difficult to accurately express a certain semantic, thereby causing token ambiguity issue. To this end, we propose a novel Speculative Jacobi Decoding with Semantics Verification (called SJD-SV), for accelerating autoregressive image generation. The key idea is that leveraging the strong correction characters between tokens to recognize semantic-aware token subsequence and then instead of perform token-by-token verification, turning to perform verification on semantic-aware token subsequence level for accelerating image generation. In particular, our method is plug-in, which can be directly integrated into existing SJD and its variants. Extensive experiments on various datasets show that existing SJD methods achieve significant performance improvement after integrating our SJD-SV method.
Vision-Language Models (VLMs) are reshaping computer vision by aligning visual and textual embeddings, allowing models to recognize visual concepts and reason about them using natural language. Traditional 3D deep-learning models, however, are typically trained for specific tasks, such as classification, segmentation, or detection, and do not naturally support cross-modal retrieval from their embedding spaces using text or images as queries. To address this issue, Contrastive Language-Image Pretraining (CLIP)-based methods align 3D embeddings with pretrained image and text representations, giving rise to 3D Vision-Language Models (3D VLMs) that support zero-shot classification, cross-modal retrieval, and open-vocabulary recognition of 3D shapes. This tutorial provides an overview of 3D VLMs, ranging from basic definitions of 3D representations and their encoding into embeddings to cross-modal contrastive alignment, modern multimodal frameworks, and 3D Vision-Large Language Models (3D VLLMs). We present the main definitions of contrastive learning for multimodal embedding alignment and highlight recent advances in language-guided 3D Gaussian splatting, 3D shape generation, and embodied AI for robotics.
展开完整摘要收起摘要↓
Vision-Language Models (VLMs) are reshaping computer vision by aligning visual and textual embeddings, allowing models to recognize visual concepts and reason about them using natural language. Traditional 3D deep-learning models, however, are typically trained for specific tasks, such as classification, segmentation, or detection, and do not naturally support cross-modal retrieval from their embedding spaces using text or images as queries. To address this issue, Contrastive Language-Image Pretraining (CLIP)-based methods align 3D embeddings with pretrained image and text representations, giving rise to 3D Vision-Language Models (3D VLMs) that support zero-shot classification, cross-modal retrieval, and open-vocabulary recognition of 3D shapes. This tutorial provides an overview of 3D VLMs, ranging from basic definitions of 3D representations and their encoding into embeddings to cross-modal contrastive alignment, modern multimodal frameworks, and 3D Vision-Large Language Models (3D VLLMs). We present the main definitions of contrastive learning for multimodal embedding alignment and highlight recent advances in language-guided 3D Gaussian splatting, 3D shape generation, and embodied AI for robotics.
Single-cell RNA sequencing studies routinely use clustering to define putative cell types and cell states, yet the observed separation may arise from sampling variability rather than genuine biological heterogeneity. This paper studies formal significance testing of such clustering structure in high-dimensional data. Existing SigClust methods assess clustering significance through Monte Carlo simulation under a Gaussian single-cluster null, but this assumption can be unreliable for normalized gene expression data and other non-Gaussian settings. We propose SigClust-LCP, a nonparametric extension that models a single cluster by a log-concave distribution. To make this approach computationally feasible in moderate to high dimensions, we develop a score-matching estimator for log-concave projection inspired by recent generative modeling ideas. We establish theoretical guarantees for the estimator and for its use in clustering significance testing. Simulations show that SigClust-LCP controls Type-I error more reliably than existing methods across a range of unimodal and mixture distributions while retaining competitive power. In a single-cell RNA sequencing analysis of Hydra cells, the method avoids spurious subclusters within annotated cell populations and supports biologically meaningful separation across lineages and body-axis regions.
展开完整摘要收起摘要↓
Single-cell RNA sequencing studies routinely use clustering to define putative cell types and cell states, yet the observed separation may arise from sampling variability rather than genuine biological heterogeneity. This paper studies formal significance testing of such clustering structure in high-dimensional data. Existing SigClust methods assess clustering significance through Monte Carlo simulation under a Gaussian single-cluster null, but this assumption can be unreliable for normalized gene expression data and other non-Gaussian settings. We propose SigClust-LCP, a nonparametric extension that models a single cluster by a log-concave distribution. To make this approach computationally feasible in moderate to high dimensions, we develop a score-matching estimator for log-concave projection inspired by recent generative modeling ideas. We establish theoretical guarantees for the estimator and for its use in clustering significance testing. Simulations show that SigClust-LCP controls Type-I error more reliably than existing methods across a range of unimodal and mixture distributions while retaining competitive power. In a single-cell RNA sequencing analysis of Hydra cells, the method avoids spurious subclusters within annotated cell populations and supports biologically meaningful separation across lineages and body-axis regions.
We study variance-preserving diffusion of the response in mixed linear regression (MLR) with unknown mixing weights. Our analysis separates the statistical guarantees of score matching from the loss geometry and optimization signal at a fixed diffusion noise level. The KL divergence links the denoising score matching objective integrated over the diffusion path with the likelihood and a terminal discrepancy. Under mild regularity conditions and terminal schedule, the resulting estimator converges up to the ground truth parameters of MLR, and its scaled error converges to the Gaussian limit of the maximum-likelihood estimator. At a fixed scale of the diffusion noise level, we derive a decomposition linking the score matching loss to cross-entropy and Expectation-Maximization (EM) operators. This decomposition yields an EM-related low-noise gradient expansion with additional correction terms of latent variance. In the high-noise limit, we further characterize gradient descent on this limiting loss under isotropic covariance. Along fixed high signal-to-noise ratio rays, the score matching imbalance gradient and the latent-variance term tend to zero pointwise. Numerical experiments illustrate our theoretical findings and statistical guarantees.
展开完整摘要收起摘要↓
We study variance-preserving diffusion of the response in mixed linear regression (MLR) with unknown mixing weights. Our analysis separates the statistical guarantees of score matching from the loss geometry and optimization signal at a fixed diffusion noise level. The KL divergence links the denoising score matching objective integrated over the diffusion path with the likelihood and a terminal discrepancy. Under mild regularity conditions and terminal schedule, the resulting estimator converges up to the ground truth parameters of MLR, and its scaled error converges to the Gaussian limit of the maximum-likelihood estimator. At a fixed scale of the diffusion noise level, we derive a decomposition linking the score matching loss to cross-entropy and Expectation-Maximization (EM) operators. This decomposition yields an EM-related low-noise gradient expansion with additional correction terms of latent variance. In the high-noise limit, we further characterize gradient descent on this limiting loss under isotropic covariance. Along fixed high signal-to-noise ratio rays, the score matching imbalance gradient and the latent-variance term tend to zero pointwise. Numerical experiments illustrate our theoretical findings and statistical guarantees.
作者Jiangyun Zhang, Kristen Surrao, Torpong Nitayanont, Yupei Zhang, Roopali Singh, Zhiyu Chen, Julia Huang, Zhou Tang, Shayan Ali Akbar, Omar Alonso, Erwin Cornejo, Yuan Li, Yi Zhang
Evaluating enterprise agents on domain-specific benchmarks is critical, yet public benchmarks rarely evaluate whether agents can integrate business knowledge with analytical computation, and constructing such benchmarks manually is costly. We present DI-Bench, a pipeline for generating realistic benchmarks for data intelligence (DI), the practice of extracting insights from large volumes of enterprise data. To emulate realistic DI tasks that require both computation and knowledge retrieval, DI-Bench builds an artifact linkage graph over data tables, dimensions, metrics, and documents to form questions involving structured data and associated knowledge. Ground truth answers are derived via query execution, followed by LLM question generation and validation. Applied to two public datasets, the pipeline produces a 731-task benchmark covering knowledge retrieval, analytical computation, and rule-grounded reasoning. To show the discriminatory capability and difficulty of the benchmark, we evaluate four models, revealing a substantial finding: models achieve only 32% accuracy when doing computational tasks where retrieved business rules modify the computation.
展开完整摘要收起摘要↓
Evaluating enterprise agents on domain-specific benchmarks is critical, yet public benchmarks rarely evaluate whether agents can integrate business knowledge with analytical computation, and constructing such benchmarks manually is costly. We present DI-Bench, a pipeline for generating realistic benchmarks for data intelligence (DI), the practice of extracting insights from large volumes of enterprise data. To emulate realistic DI tasks that require both computation and knowledge retrieval, DI-Bench builds an artifact linkage graph over data tables, dimensions, metrics, and documents to form questions involving structured data and associated knowledge. Ground truth answers are derived via query execution, followed by LLM question generation and validation. Applied to two public datasets, the pipeline produces a 731-task benchmark covering knowledge retrieval, analytical computation, and rule-grounded reasoning. To show the discriminatory capability and difficulty of the benchmark, we evaluate four models, revealing a substantial finding: models achieve only 32% accuracy when doing computational tasks where retrieved business rules modify the computation.
Lifelong LLM agents increasingly rely on external skill artifacts as one element for preserving and reusing capabilities over time. These skills (usually portable Markdown files such as SKILL.md) describe when and how to apply a capability and must be corrected, expanded, and consolidated as tools and usage patterns shift over deployment. Recent work seeks to automate skill curation, but it largely evaluates against automated baselines and treats human maintenance as an unmeasured bottleneck. We study that missing process directly. We mine the full commit histories of five public AI-skill repositories, a purposive sample of AI-tooling organizations, covering 873 commits, 143 skill files, and 254 substantive post-creation edits from October 2025 to June 2026. We code each edit with pre-registered governance, operation, and trigger-evidence codebooks. Three findings emerge. First, every substantive edit is authored or merged through a named human account, while 62% carry an AI co-author trailer, with large repository-level variation. Second, these edits are genuine curation: an audited sample shows that most change skill content, and the coded operations are dominated by additions and corrections. Third, a pre-registered rule-likeness axis fails its reliability gate; reliably coding rule-likeness from commit artifacts remains an open measurement problem. We release the corpus, codebooks, mining scripts, and a replay protocol for automated skill curators. For self-evolving agents, public skill maintenance currently looks less like an autonomous pipeline than a human-governed, AI-assisted loop that future curators must measure against and operate within.
展开完整摘要收起摘要↓
Lifelong LLM agents increasingly rely on external skill artifacts as one element for preserving and reusing capabilities over time. These skills (usually portable Markdown files such as SKILL.md) describe when and how to apply a capability and must be corrected, expanded, and consolidated as tools and usage patterns shift over deployment. Recent work seeks to automate skill curation, but it largely evaluates against automated baselines and treats human maintenance as an unmeasured bottleneck. We study that missing process directly. We mine the full commit histories of five public AI-skill repositories, a purposive sample of AI-tooling organizations, covering 873 commits, 143 skill files, and 254 substantive post-creation edits from October 2025 to June 2026. We code each edit with pre-registered governance, operation, and trigger-evidence codebooks. Three findings emerge. First, every substantive edit is authored or merged through a named human account, while 62% carry an AI co-author trailer, with large repository-level variation. Second, these edits are genuine curation: an audited sample shows that most change skill content, and the coded operations are dominated by additions and corrections. Third, a pre-registered rule-likeness axis fails its reliability gate; reliably coding rule-likeness from commit artifacts remains an open measurement problem. We release the corpus, codebooks, mining scripts, and a replay protocol for automated skill curators. For self-evolving agents, public skill maintenance currently looks less like an autonomous pipeline than a human-governed, AI-assisted loop that future curators must measure against and operate within.
作者Hoyeol Yang, Woojung Song, Taewon Kim, Jonghyun Song, Seoyeon Park, Yohan Jo
Existing evaluations of tool-using agents primarily measure whether an agent can successfully complete diverse tasks with tools. These evaluations generally assume that tools return reliable information. However, tool returns in real-world systems can be plausible yet incorrect. We investigate how agents respond to unreliable tool returns by evaluating fourteen LLMs using three tools-web search, LLM sub-agent delegation, and code execution. For each tool, we corrupt its returns and measure whether agents adopt the corrupted content in their final answers. Agents exhibit high levels of overtrust across all three settings: the mean adoption rate exceeds one third for every tool and reaches 68.0% for web search. Analysis of reasoning traces reveals a particularly concerning failure mode: agents often recognize conflicts and even recover the correct answer internally, yet present only the corrupted answer without warning the user. To mitigate agents' overtrust in tool returns, we intervene at three levels: prompting by the user, metadata from the tool provider, and post-training by the agent builder. Although some interventions help for particular models or tools, none consistently mitigates overtrust across tools. These findings identify overtrust in unreliable tools as a serious and persistent failure mode, motivating evaluations and interventions that enable agents to validate tool outputs and transparently communicate unresolved conflicts.
展开完整摘要收起摘要↓
Existing evaluations of tool-using agents primarily measure whether an agent can successfully complete diverse tasks with tools. These evaluations generally assume that tools return reliable information. However, tool returns in real-world systems can be plausible yet incorrect. We investigate how agents respond to unreliable tool returns by evaluating fourteen LLMs using three tools-web search, LLM sub-agent delegation, and code execution. For each tool, we corrupt its returns and measure whether agents adopt the corrupted content in their final answers. Agents exhibit high levels of overtrust across all three settings: the mean adoption rate exceeds one third for every tool and reaches 68.0% for web search. Analysis of reasoning traces reveals a particularly concerning failure mode: agents often recognize conflicts and even recover the correct answer internally, yet present only the corrupted answer without warning the user. To mitigate agents' overtrust in tool returns, we intervene at three levels: prompting by the user, metadata from the tool provider, and post-training by the agent builder. Although some interventions help for particular models or tools, none consistently mitigates overtrust across tools. These findings identify overtrust in unreliable tools as a serious and persistent failure mode, motivating evaluations and interventions that enable agents to validate tool outputs and transparently communicate unresolved conflicts.
作者T. J. Barton, Chris Constantakis, Patti Hauseman, Annie Mous, Alaska Hoffman, Brian Bergeron, Hunter Goodreau
We present a continuous, population-scale measurement record of autonomous language-model trading agents operating in production across two systems with one design lineage: DX Terminal Pro (3,505 user-funded vaults trading real ETH in Base memecoin markets for 21 days, February to March 2026) and the DXAP live alpha fleet (500 to 599 user-created agents all-history, 91 to 117 concurrently active, trading Hyperliquid perpetuals, June to August 2026). The record spans roughly six months, 7.5M single-model invocations with about 300K onchain actions, and a further 231,638 multi-tool turns producing 14,596 fills. Four findings carry the paper. First, the operating layer determines behavior more than anything written in strategy text: a risk slider explains leverage (+0.425 per level), agent fixed effects absorb 60% of variance, and a leaderboard render boundary causally routes selection (regression discontinuity 1.75x at the top-3 cut). Second, sizing is volatility-blind: median leverage is 5.0x in every volatility sextile, and one posture-slider cell (11% of the book) holds 62% of liquidations. Third, agents capture almost none of the upside they reach: 43.2% of positions saw at least +300 bps of favorable excursion within 24h, yet 49.3% of those closed with a negative trade return; a mechanical bracket recovers +39.0 bps per position. Fourth, neither fleet shows a directional edge. The DXAP fleet is not profitable and trails a matched Hyperliquid retail benchmark (41% vs. 50% roundtrip win rate). A paired-replay league of frontier models on 416 captured production scenarios finds decision quality statistically indistinguishable at this horizon, while choice stability differs sharply across model families. Every headline survives day-clustered inference, permutation nulls, and a common-fee restatement; the paper closes with a 17-rule methodology canon bought with our own retractions.
展开完整摘要收起摘要↓
We present a continuous, population-scale measurement record of autonomous language-model trading agents operating in production across two systems with one design lineage: DX Terminal Pro (3,505 user-funded vaults trading real ETH in Base memecoin markets for 21 days, February to March 2026) and the DXAP live alpha fleet (500 to 599 user-created agents all-history, 91 to 117 concurrently active, trading Hyperliquid perpetuals, June to August 2026). The record spans roughly six months, 7.5M single-model invocations with about 300K onchain actions, and a further 231,638 multi-tool turns producing 14,596 fills. Four findings carry the paper. First, the operating layer determines behavior more than anything written in strategy text: a risk slider explains leverage (+0.425 per level), agent fixed effects absorb 60% of variance, and a leaderboard render boundary causally routes selection (regression discontinuity 1.75x at the top-3 cut). Second, sizing is volatility-blind: median leverage is 5.0x in every volatility sextile, and one posture-slider cell (11% of the book) holds 62% of liquidations. Third, agents capture almost none of the upside they reach: 43.2% of positions saw at least +300 bps of favorable excursion within 24h, yet 49.3% of those closed with a negative trade return; a mechanical bracket recovers +39.0 bps per position. Fourth, neither fleet shows a directional edge. The DXAP fleet is not profitable and trails a matched Hyperliquid retail benchmark (41% vs. 50% roundtrip win rate). A paired-replay league of frontier models on 416 captured production scenarios finds decision quality statistically indistinguishable at this horizon, while choice stability differs sharply across model families. Every headline survives day-clustered inference, permutation nulls, and a common-fee restatement; the paper closes with a 17-rule methodology canon bought with our own retractions.
The paradigm of LLMs has rapidly shifted from passive language interfaces to autonomous Claw-like agents that execute long-horizon tasks across stateful workspaces. While Agentic Reinforcement Learning (Agentic RL) provides a promising path to optimize these agents, its scaling is heavily bottlenecked by the severe scarcity of interactive training environments. Existing synthetic environments are strictly limited to tool-calling endpoints, rendering them insufficient for accommodating the end-to-end real-world demands of claw-like agents. To bridge this gap, we introduce EnvCraft, an automated framework for synthesizing executable environments and scalable training data. Specifically, EnvCraft employs an environment synthesis engine to build sandbox-isolated workspaces, alongside a topology-aware data generation engine to produce coherent task trajectories. Overall, we synthesize 139 interactive environments comprising approximately 20K complex tasks for Agentic RL training. Experiments on Qwen3/3.5 models (8B-32B) show that our method yields gains of up to +11.9% on Claw-style benchmarks and +8.0% on general tool-use benchmarks, with concurrent reductions in inference token cost. The results confirm that synthesized executable environments provide robust and generalizable learning signals for training.
展开完整摘要收起摘要↓
The paradigm of LLMs has rapidly shifted from passive language interfaces to autonomous Claw-like agents that execute long-horizon tasks across stateful workspaces. While Agentic Reinforcement Learning (Agentic RL) provides a promising path to optimize these agents, its scaling is heavily bottlenecked by the severe scarcity of interactive training environments. Existing synthetic environments are strictly limited to tool-calling endpoints, rendering them insufficient for accommodating the end-to-end real-world demands of claw-like agents. To bridge this gap, we introduce EnvCraft, an automated framework for synthesizing executable environments and scalable training data. Specifically, EnvCraft employs an environment synthesis engine to build sandbox-isolated workspaces, alongside a topology-aware data generation engine to produce coherent task trajectories. Overall, we synthesize 139 interactive environments comprising approximately 20K complex tasks for Agentic RL training. Experiments on Qwen3/3.5 models (8B-32B) show that our method yields gains of up to +11.9% on Claw-style benchmarks and +8.0% on general tool-use benchmarks, with concurrent reductions in inference token cost. The results confirm that synthesized executable environments provide robust and generalizable learning signals for training.
作者Katherine Tieu, Dongqi Fu, Yinglong Xia, Hong Li, Hong Yan, Jingrui He
Recent multi-agent LLM systems increasingly rely on graph-structured communication to coordinate specialized agents. We revisit multi-agent orchestration from a graph-engineering perspective: rather than optimizing a static topology, we synthesize a task-conditioned temporal workflow graph that jointly specifies agent connectivity and edge-level communication semantics. We introduce ReActNet, a training-free framework that compiles a query and a set of role-specialized agents into a sequence of directed communication graphs. Each graph snapshot corresponds to one reasoning stage, and each edge carries a natural-language instruction specifying the message that a source agent should provide to a target agent. The compiled temporal graph is then executed through structured message passing: agents update their reasoning states by integrating their previous states with messages from controller-assigned neighbors, and a final aggregator synthesizes the resulting states into the answer. This design separates graph compilation from graph execution, making multi-agent coordination explicit, inspectable, and task-conditioned without requiring reinforcement learning or gradient-based topology optimization. Across knowledge reasoning, mathematical problem solving, code generation, and GAIA-style assistant tasks, ReActNet consistently improves over fixed-topology and learned-topology baselines while maintaining competitive inference cost. These results suggest that effective multi-agent orchestration depends not only on which agents communicate, but also on engineering executable workflow graphs that encode when, why, and how information should flow during reasoning.
展开完整摘要收起摘要↓
Recent multi-agent LLM systems increasingly rely on graph-structured communication to coordinate specialized agents. We revisit multi-agent orchestration from a graph-engineering perspective: rather than optimizing a static topology, we synthesize a task-conditioned temporal workflow graph that jointly specifies agent connectivity and edge-level communication semantics. We introduce ReActNet, a training-free framework that compiles a query and a set of role-specialized agents into a sequence of directed communication graphs. Each graph snapshot corresponds to one reasoning stage, and each edge carries a natural-language instruction specifying the message that a source agent should provide to a target agent. The compiled temporal graph is then executed through structured message passing: agents update their reasoning states by integrating their previous states with messages from controller-assigned neighbors, and a final aggregator synthesizes the resulting states into the answer. This design separates graph compilation from graph execution, making multi-agent coordination explicit, inspectable, and task-conditioned without requiring reinforcement learning or gradient-based topology optimization. Across knowledge reasoning, mathematical problem solving, code generation, and GAIA-style assistant tasks, ReActNet consistently improves over fixed-topology and learned-topology baselines while maintaining competitive inference cost. These results suggest that effective multi-agent orchestration depends not only on which agents communicate, but also on engineering executable workflow graphs that encode when, why, and how information should flow during reasoning.
作者Piyush Sao, Narasinga Miniskar, Pedro Valero-Lara, Keita Teranishi, Sudip Seal
We consider the problem of minimizing error in quantized matrix multiplication $C=AB$. Scalar quantization of the factors introduces rounding errors whose scale depends on the maximum absolute entries — the ranges — of their rows and columns. These ranges determine the quantization grid steps. To reduce the error, we optimize over product-preserving transformations that alter the factor ranges and grid steps without changing $C$. Specifically, we seek the smallest leading expected squared error over invertible inner changes of basis and orthogonal outer rotations. Under independent, zero-mean subtractive dither noise on an unbounded lattice, we prove the output-only bound $E_{\rm lead} \ge (c_A+c_B)/K \Vert AB\Vert_*^2$, where $K$ is the inner dimension, $c_A$ and $c_B$ are normalized noise variances, and $\Vert AB\Vert_*$ is the nuclear norm. The bound is tight: an SVD-aligned Hadamard construction attains the infimum whenever a Hadamard matrix of order $K$ exists, including every power of two, while an SVD-aligned DCT construction is within a factor of two for every $K$. Without outer rotations, Gram-matrix balancing minimizes factorization energy, and finite-set flattening achieves the bound within $C\log(K(m+n))$. For power-of-two $K$, conditional expectations deterministically select the Hadamard signs in $O((m+n)K^2)$ exact-real operations. Synthetic experiments verify both constructions and illustrate the tradeoff between regularization and conditioning. These results characterize the full-gauge optimum and quantify the cost of preserving row and column indices.
展开完整摘要收起摘要↓
We consider the problem of minimizing error in quantized matrix multiplication $C=AB$. Scalar quantization of the factors introduces rounding errors whose scale depends on the maximum absolute entries — the ranges — of their rows and columns. These ranges determine the quantization grid steps. To reduce the error, we optimize over product-preserving transformations that alter the factor ranges and grid steps without changing $C$. Specifically, we seek the smallest leading expected squared error over invertible inner changes of basis and orthogonal outer rotations. Under independent, zero-mean subtractive dither noise on an unbounded lattice, we prove the output-only bound $E_{\rm lead} \ge (c_A+c_B)/K \Vert AB\Vert_*^2$, where $K$ is the inner dimension, $c_A$ and $c_B$ are normalized noise variances, and $\Vert AB\Vert_*$ is the nuclear norm. The bound is tight: an SVD-aligned Hadamard construction attains the infimum whenever a Hadamard matrix of order $K$ exists, including every power of two, while an SVD-aligned DCT construction is within a factor of two for every $K$. Without outer rotations, Gram-matrix balancing minimizes factorization energy, and finite-set flattening achieves the bound within $C\log(K(m+n))$. For power-of-two $K$, conditional expectations deterministically select the Hadamard signs in $O((m+n)K^2)$ exact-real operations. Synthetic experiments verify both constructions and illustrate the tradeoff between regularization and conditioning. These results characterize the full-gauge optimum and quantify the cost of preserving row and column indices.
作者Cen Mia Zhao, Peng Wang, Chuan Shi, Yufeng Zhang, Ying Lyu, Wanmeng Ren, Robert Xue, Claire Na Cheng, Yashar Mehdad
Conversational assistants can blend retrieval, action selection, escalation, and wording in a single model path, or separate those roles. We report a production migration of a customer-support assistant at a large accommodation marketplace (millions of conversations per month, 11 languages, 10-second P90). Dynamic Response (DR) replaces a single Qwen3-235B-A22B blended responder with a bounded ReAct orchestrator over typed tools plus a smaller generator that writes from a backend-validated context contract. Because the migration also changed prompts, alignment, and serving, we attribute each effect to its cause and claim as architecture effects only those measured on identical replayed turns: typed entity selection moves the reservation selector to a precision-first operating point (precision 8.3% to 89.1%, recall 75.2% to 67.3%), and typed action IDs with a membership check remove observed structured-action hallucination (2.14% to 0.0%). A low-ramp A/B test reproduces the replay escalation reductions: hard-escalation responses fall from 5.60% to 3.08% and soft-escalation responses from 9.56% to 2.49%, while production handoff volume holds roughly steady; self-solve is directional (+5.1 points, 95% CI [-2, +12]). Serving optimizations cut orchestrator P90 latency from 3.87s to 2.24s on a GPU footprint reduced by roughly one-third, and self-hosting reduces estimated annual model-serving cost by more than an order of magnitude.
展开完整摘要收起摘要↓
Conversational assistants can blend retrieval, action selection, escalation, and wording in a single model path, or separate those roles. We report a production migration of a customer-support assistant at a large accommodation marketplace (millions of conversations per month, 11 languages, 10-second P90). Dynamic Response (DR) replaces a single Qwen3-235B-A22B blended responder with a bounded ReAct orchestrator over typed tools plus a smaller generator that writes from a backend-validated context contract. Because the migration also changed prompts, alignment, and serving, we attribute each effect to its cause and claim as architecture effects only those measured on identical replayed turns: typed entity selection moves the reservation selector to a precision-first operating point (precision 8.3% to 89.1%, recall 75.2% to 67.3%), and typed action IDs with a membership check remove observed structured-action hallucination (2.14% to 0.0%). A low-ramp A/B test reproduces the replay escalation reductions: hard-escalation responses fall from 5.60% to 3.08% and soft-escalation responses from 9.56% to 2.49%, while production handoff volume holds roughly steady; self-solve is directional (+5.1 points, 95% CI [-2, +12]). Serving optimizations cut orchestrator P90 latency from 3.87s to 2.24s on a GPU footprint reduced by roughly one-third, and self-hosting reduces estimated annual model-serving cost by more than an order of magnitude.
Personal LLM assistants (health companions, elder-care agents, accessibility aides) are judged by what they remember about a person: a medication or an allergy mentioned in passing and needed days later. Privacy pushes them on-device, where a month of conversation can outgrow the model's own weights, so an eviction policy must decide what the cache forgets. Benchmarks report that eviction keeps such facts at a 20% budget, but they compress a prompt that already contains the user's future question, foresight no cache-reusing assistant has. Hide the question until after compression and the advantage vanishes: on PA-Bench, 100 assistant conversations we construct, an allergy mentioned in passing survives to the question that needs it 0--1% of the time, against 97% with full memory. The cause is the budget, not the scorer: none of the training-free policies we evaluate ranks the fact high enough, and the budget that would keep it is too large to bother compressing. A compressed cache is an inference-reuse mechanism, not a persistence layer: safety-critical facts need an auditable episodic store alongside it, and an interface that asks rather than invents.
展开完整摘要收起摘要↓
Personal LLM assistants (health companions, elder-care agents, accessibility aides) are judged by what they remember about a person: a medication or an allergy mentioned in passing and needed days later. Privacy pushes them on-device, where a month of conversation can outgrow the model's own weights, so an eviction policy must decide what the cache forgets. Benchmarks report that eviction keeps such facts at a 20% budget, but they compress a prompt that already contains the user's future question, foresight no cache-reusing assistant has. Hide the question until after compression and the advantage vanishes: on PA-Bench, 100 assistant conversations we construct, an allergy mentioned in passing survives to the question that needs it 0--1% of the time, against 97% with full memory. The cause is the budget, not the scorer: none of the training-free policies we evaluate ranks the fact high enough, and the budget that would keep it is too large to bother compressing. A compressed cache is an inference-reuse mechanism, not a persistence layer: safety-critical facts need an auditable episodic store alongside it, and an interface that asks rather than invents.
作者Zlatan Feric, Amir Taherin, Bin Ren, Yanzhi Wang, Jennifer Dy, David Kaeli
We present RAGMark, a modular benchmarking framework for advanced Retrieval-Augmented Generation (RAG) systems targeting small-scale multi-GPU environments. RAGMark evaluates diverse RAG components, including retrievers, vector databases, prompt-processing methods, and generator models, while collecting detailed per-stage metrics such as latency, GPU utilization, memory consumption, power usage, time to first token (TTFT), throughput, and answer quality. The framework is highly extensible, separating RAG stages, timing, and resource monitoring into modular components, and is designed to efficiently sweep large configuration spaces while minimizing repeated model and database initialization overhead. Using RAGMark, we characterize five RAG workloads on open-domain QA datasets across varying retrieval depths, model scales, reranking, compression methods, and vector database configurations. We show that while autoregressive generation dominates latency in naive pipelines, context-reduction techniques shift bottlenecks across compute, memory bandwidth, and preprocessing stages. Reranking and compression produce compounding benefits: reranking reduces compression workload itself, while both jointly reduce prefill and KV-cache traversal costs, lowering energy consumption by up to 66%. We further observe strong cross-stage interactions, where small upstream context reductions cascade through downstream latency, memory traffic, and energy consumption. The RAGMark source code is publicly available at: https://github.com/zferic/RAGMark.
展开完整摘要收起摘要↓
We present RAGMark, a modular benchmarking framework for advanced Retrieval-Augmented Generation (RAG) systems targeting small-scale multi-GPU environments. RAGMark evaluates diverse RAG components, including retrievers, vector databases, prompt-processing methods, and generator models, while collecting detailed per-stage metrics such as latency, GPU utilization, memory consumption, power usage, time to first token (TTFT), throughput, and answer quality. The framework is highly extensible, separating RAG stages, timing, and resource monitoring into modular components, and is designed to efficiently sweep large configuration spaces while minimizing repeated model and database initialization overhead. Using RAGMark, we characterize five RAG workloads on open-domain QA datasets across varying retrieval depths, model scales, reranking, compression methods, and vector database configurations. We show that while autoregressive generation dominates latency in naive pipelines, context-reduction techniques shift bottlenecks across compute, memory bandwidth, and preprocessing stages. Reranking and compression produce compounding benefits: reranking reduces compression workload itself, while both jointly reduce prefill and KV-cache traversal costs, lowering energy consumption by up to 66%. We further observe strong cross-stage interactions, where small upstream context reductions cascade through downstream latency, memory traffic, and energy consumption. The RAGMark source code is publicly available at: https://github.com/zferic/RAGMark.
作者Jiahao Zheng, Yifan Qin, Xiaobo Sharon Hu, Yiyu Shi
The key-value (KV) cache is the dominant memory bottleneck in long-context large language model (LLM) decoding: every step reads it entirely, so decoding is memory-bandwidth bound. Holding a quantized KV cache in dense on-chip non-volatile memory (NVM) removes the off-chip transfer. Existing KV quantization methods, however, were designed for GPU-style memory systems: KIVI attaches per-group metadata, adding about 25% to the stored KV cache; KVQuant keeps sparse full-precision outliers that a dense array cannot hold in place. This paper examines what these structures cost when the KV cache resides in NVM behind fixed-range converters, and designs a quantization scheme matched to that interface. The architecture stores the quantized KV cache in dense on-chip NVM, uses a small static analog crossbar only for the fixed rotation, and keeps attention in on-chip digital logic. A randomized rotation and per-vector normalization give every coordinate the same range, so one fixed codebook for keys and one for values, each shared across all tokens of the corresponding tensor type, serve the entire KV cache. The codebook thresholds are programmed once as the read converter's reference levels, enabling fixed-range digitization with no per-token converter reconfiguration. Dequantization is a sixteen-entry lookup and one norm multiply; the only per-vector metadata is one scalar, about 3%. Across models from 3B to 14B and contexts to 32k tokens, the four-bit KV cache maintains accuracy under storage and crossbar noise simulated at realistic device levels. KIVI and KVQuant remain more accurate in software; the advantage of our format lies at the memory interface: 3.1-3.6x lower KV read energy than both mapped to the same NVM, and 8x lower metadata overhead than KIVI. The contribution is a KV quantization co-designed with the NVM memory interface rather than a new accuracy record.
展开完整摘要收起摘要↓
The key-value (KV) cache is the dominant memory bottleneck in long-context large language model (LLM) decoding: every step reads it entirely, so decoding is memory-bandwidth bound. Holding a quantized KV cache in dense on-chip non-volatile memory (NVM) removes the off-chip transfer. Existing KV quantization methods, however, were designed for GPU-style memory systems: KIVI attaches per-group metadata, adding about 25% to the stored KV cache; KVQuant keeps sparse full-precision outliers that a dense array cannot hold in place. This paper examines what these structures cost when the KV cache resides in NVM behind fixed-range converters, and designs a quantization scheme matched to that interface. The architecture stores the quantized KV cache in dense on-chip NVM, uses a small static analog crossbar only for the fixed rotation, and keeps attention in on-chip digital logic. A randomized rotation and per-vector normalization give every coordinate the same range, so one fixed codebook for keys and one for values, each shared across all tokens of the corresponding tensor type, serve the entire KV cache. The codebook thresholds are programmed once as the read converter's reference levels, enabling fixed-range digitization with no per-token converter reconfiguration. Dequantization is a sixteen-entry lookup and one norm multiply; the only per-vector metadata is one scalar, about 3%. Across models from 3B to 14B and contexts to 32k tokens, the four-bit KV cache maintains accuracy under storage and crossbar noise simulated at realistic device levels. KIVI and KVQuant remain more accurate in software; the advantage of our format lies at the memory interface: 3.1-3.6x lower KV read energy than both mapped to the same NVM, and 8x lower metadata overhead than KIVI. The contribution is a KV quantization co-designed with the NVM memory interface rather than a new accuracy record.
作者Esteban Feuerstein, Victoria Klimkowski, Juan Manuel Ortiz de Zarate, Federico Hernán Suaiter
Understanding whether language-model embeddings encode structured real-world information is important for both representation analysis and information retrieval. We study this question for temporal and geographic signals using a simple projection-based method that operates directly on output embeddings. Given a small set of seed examples, the method defines an axis in embedding space and ranks texts or entities by their projection onto that axis. Our approach is fully black-box and model-agnostic: it requires only embeddings, without access to model weights, internal activations, auxiliary probes, or additional training. This makes it applicable to modern embedding models available only through APIs and provides a lightweight way to analyze whether temporal and spatial dimensions are present in their representation spaces. We apply the method to temporal and geographic datasets and find that embedding projections recover meaningful chronological and spatial structure. These results provide evidence that output embeddings encode signals relevant to time and space, while also offering a practical tool for interpretability and for downstream temporal and geographic information retrieval tasks, such as temporal ordering, geographic ranking, and tagging.
展开完整摘要收起摘要↓
Understanding whether language-model embeddings encode structured real-world information is important for both representation analysis and information retrieval. We study this question for temporal and geographic signals using a simple projection-based method that operates directly on output embeddings. Given a small set of seed examples, the method defines an axis in embedding space and ranks texts or entities by their projection onto that axis. Our approach is fully black-box and model-agnostic: it requires only embeddings, without access to model weights, internal activations, auxiliary probes, or additional training. This makes it applicable to modern embedding models available only through APIs and provides a lightweight way to analyze whether temporal and spatial dimensions are present in their representation spaces. We apply the method to temporal and geographic datasets and find that embedding projections recover meaningful chronological and spatial structure. These results provide evidence that output embeddings encode signals relevant to time and space, while also offering a practical tool for interpretability and for downstream temporal and geographic information retrieval tasks, such as temporal ordering, geographic ranking, and tagging.
In common-context question answering (CCQA) tasks, multiple input questions share a common context to base their answers from. However, Large Language Models typically generate each answer using an independent prompt. While existing batching and caching techniques help improve parallelism and reduce repeated computations, the separation of questions across prompts limits the achievable speedup, as modern GPUs are underutilized due to a memory bottleneck during attention. We present Intra-Prompt Parallel Decoding (IPPD), a novel inference method that answers multiple common-context questions in parallel within a single prompt. IPPD directly addresses the bottleneck by efficiently sharing both memory and computation during the attention process, as the next token for every question is decoded in a single inference step. IPPD uses virtual position IDs and attention mask manipulation to generate the same output as standard prompting without requiring fine-tuning or any changes to the LLM architecture. Since all parallelism occurs within a prompt, IPPD is fully compatible with batched inference, even when each prompt features a different context. Our experiments show that IPPD delivers up to 7X the effective throughput as standard decoding without quality degradation, and outperforms prefix caching with PagedAttention in most settings.
展开完整摘要收起摘要↓
In common-context question answering (CCQA) tasks, multiple input questions share a common context to base their answers from. However, Large Language Models typically generate each answer using an independent prompt. While existing batching and caching techniques help improve parallelism and reduce repeated computations, the separation of questions across prompts limits the achievable speedup, as modern GPUs are underutilized due to a memory bottleneck during attention. We present Intra-Prompt Parallel Decoding (IPPD), a novel inference method that answers multiple common-context questions in parallel within a single prompt. IPPD directly addresses the bottleneck by efficiently sharing both memory and computation during the attention process, as the next token for every question is decoded in a single inference step. IPPD uses virtual position IDs and attention mask manipulation to generate the same output as standard prompting without requiring fine-tuning or any changes to the LLM architecture. Since all parallelism occurs within a prompt, IPPD is fully compatible with batched inference, even when each prompt features a different context. Our experiments show that IPPD delivers up to 7X the effective throughput as standard decoding without quality degradation, and outperforms prefix caching with PagedAttention in most settings.
We identify a special group of token vectors inside large language models (LLMs), which we term magnetic vectors, that organize the surrounding tokens by either attracting or repelling them. Particularly, tokens pointing the same way as an attracting magnet are elongated; tokens pointing the same way as a repelling magnet are compressed. Just as physical magnets pull or push away the iron filings around them, these vectors organize their surroundings through two opposing polarities. Moreover, we identify a statistically significant pattern in linguistic category where function words consistently act as repelling magnets in early layers, and we also find magnets consistently reorganize their polarities in unique ways deeper in the model. In a further case study we find this observation may unveil a deliberate, layer-wise organization in how LLMs process language. This pattern is consistent across different LLM architectures, sizes, and layer configurations. It is also causally relevant. When the LLM is fine-tuned for a downstream task, the task-functional tokens emerge as magnets. E.g., in question answering, the answer-span tokens become uniquely repelling magnets in the final layer, geometrically carving the answer out of the surrounding context. Furthermore, removing early-layer repelling magnets devastates syntactic tasks (POS tagging accuracy drops from 91% to below 10%) while sparing semantic ones, and removing late-layer attracting magnets does the reverse. We believe this phenomenon warrants further investigation, as it opens the first probe-free path to understanding how language models geometrically organize linguistic computation across their layers.
展开完整摘要收起摘要↓
We identify a special group of token vectors inside large language models (LLMs), which we term magnetic vectors, that organize the surrounding tokens by either attracting or repelling them. Particularly, tokens pointing the same way as an attracting magnet are elongated; tokens pointing the same way as a repelling magnet are compressed. Just as physical magnets pull or push away the iron filings around them, these vectors organize their surroundings through two opposing polarities. Moreover, we identify a statistically significant pattern in linguistic category where function words consistently act as repelling magnets in early layers, and we also find magnets consistently reorganize their polarities in unique ways deeper in the model. In a further case study we find this observation may unveil a deliberate, layer-wise organization in how LLMs process language. This pattern is consistent across different LLM architectures, sizes, and layer configurations. It is also causally relevant. When the LLM is fine-tuned for a downstream task, the task-functional tokens emerge as magnets. E.g., in question answering, the answer-span tokens become uniquely repelling magnets in the final layer, geometrically carving the answer out of the surrounding context. Furthermore, removing early-layer repelling magnets devastates syntactic tasks (POS tagging accuracy drops from 91% to below 10%) while sparing semantic ones, and removing late-layer attracting magnets does the reverse. We believe this phenomenon warrants further investigation, as it opens the first probe-free path to understanding how language models geometrically organize linguistic computation across their layers.
作者Jianxin Gao, Tianyi Yu, Linna Deng, Runze Li, Zining Wang
Production multi-agent systems replace agents constantly, on the assumption that an agent filling a role is interchangeable with any other agent that can do the job. We test that assumption. Eight teams per setting are formed independently from one base model on the same tasks, each agent keeping a private notebook across ten formation episodes; we then trade role-matched agents between teams and measure what changes on held-out tasks. Against a placebo that reproduces the disruption of a roster change without changing who occupies the seat, a swap costs little in task score but raises the communication a team spends per unit of progress by 16 to 63 percent, and in Hanabi a swapped agent is more expensive than an inexperienced one, consistent with interference from conventions learned with its former partner. In Collab-Overcooked, when the agent that sets the agenda is replaced, most of the extra communication comes from the agent that stayed. Three ablations, over base models, decoding temperature and formation length, move the swap penalty alongside one other quantity: how far independently formed teams drift apart. Greedy decoding lowers both; doubling a team's history raises both. In these settings, agents are more fungible in task outcome than in coordination efficiency, with larger swap effects after longer formation histories.
展开完整摘要收起摘要↓
Production multi-agent systems replace agents constantly, on the assumption that an agent filling a role is interchangeable with any other agent that can do the job. We test that assumption. Eight teams per setting are formed independently from one base model on the same tasks, each agent keeping a private notebook across ten formation episodes; we then trade role-matched agents between teams and measure what changes on held-out tasks. Against a placebo that reproduces the disruption of a roster change without changing who occupies the seat, a swap costs little in task score but raises the communication a team spends per unit of progress by 16 to 63 percent, and in Hanabi a swapped agent is more expensive than an inexperienced one, consistent with interference from conventions learned with its former partner. In Collab-Overcooked, when the agent that sets the agenda is replaced, most of the extra communication comes from the agent that stayed. Three ablations, over base models, decoding temperature and formation length, move the swap penalty alongside one other quantity: how far independently formed teams drift apart. Greedy decoding lowers both; doubling a team's history raises both. In these settings, agents are more fungible in task outcome than in coordination efficiency, with larger swap effects after longer formation histories.
Large language model (LLM)-based multi-agent systems have experienced rapid growth in recent years. Despite their promise, such systems remain fragile, frequently exhibiting reasoning and coordination errors that can lead to system-level failures. Failure attribution in such systems relies on tracing natural language interactions among agents to identify the decisive error, which refers to the earliest action whose correction can reverse system failure. There are two key challenges: 1) Shallow attribution: Existing methods often capture only minor deviations, such as incomplete retrievals or formatting errors, which verification mechanisms can correct, while missing the decisive cause of system failure. 2) Contextual degradation: As the length of the system traces increases, the model's reasoning ability rapidly deteriorates. To address these challenges, we propose DCFA, a training-free framework for failure attribution. DCFA integrates a global module that constructs structured causal-inspired dependency graphs from system traces to identify the initial decisive error, and a local module that applies local counterfactual-inspired reasoning to refine causal-inspired attribution. Experiments on the Who&When benchmark across six LLMs show that DCFA improves step-level accuracy by up to 8.27% over state-of-the-art baselines.
展开完整摘要收起摘要↓
Large language model (LLM)-based multi-agent systems have experienced rapid growth in recent years. Despite their promise, such systems remain fragile, frequently exhibiting reasoning and coordination errors that can lead to system-level failures. Failure attribution in such systems relies on tracing natural language interactions among agents to identify the decisive error, which refers to the earliest action whose correction can reverse system failure. There are two key challenges: 1) Shallow attribution: Existing methods often capture only minor deviations, such as incomplete retrievals or formatting errors, which verification mechanisms can correct, while missing the decisive cause of system failure. 2) Contextual degradation: As the length of the system traces increases, the model's reasoning ability rapidly deteriorates. To address these challenges, we propose DCFA, a training-free framework for failure attribution. DCFA integrates a global module that constructs structured causal-inspired dependency graphs from system traces to identify the initial decisive error, and a local module that applies local counterfactual-inspired reasoning to refine causal-inspired attribution. Experiments on the Who&When benchmark across six LLMs show that DCFA improves step-level accuracy by up to 8.27% over state-of-the-art baselines.
Chilli (Capsicum annuum) is one of India's most economically significant crops, yet its productivity is persistently threatened by diseases that are difficult to identify without expert intervention. While Vision Transformers (ViTs) have achieved high classification accuracy, their large computational footprint makes deployment on resource constrained devices challenging. Existing compression approaches typically address pruning, quantization, and knowledge distillation in isolation, leaving the potential benefits and interactions of their combined application insufficiently explored. We propose a unified Vision Transformer compression framework that combines Hessian-Balanced Adaptive Block Pruning (H-BAC), guided by second-order sensitivity estimation, with quantization and attention-based knowledge distillation. To systematically identify the most effective configuration within each compression family, each technique is first evaluated independently through controlled ablation studies, after which the best-performing components are integrated into a sequential deployment pipeline tailored to real-world agricultural constraints. On a chilli 3-class village-split dataset with a genuine cross-village, cross-device out-of-distribution test split, the resulting compressed models match or exceed the 95.13% FP32 baseline's accuracy, alongside 74-98% model size reduction, and the fully integrated compression pipeline achieves a 54.5x size reduction (327.42 MB to 6.01 MB) at 95.13 +/- 2.32% accuracy across four tested configurations. A direct comparison further reveals that, on this dataset, a directly-trained student of the same final size, without pruning or distillation, reaches comparable accuracy of 94.87%, at the same 6.01 MB INT8 size, indicating where H-BAC and knowledge distillation are, and are not yet shown to be, worth their computational cost.
展开完整摘要收起摘要↓
Chilli (Capsicum annuum) is one of India's most economically significant crops, yet its productivity is persistently threatened by diseases that are difficult to identify without expert intervention. While Vision Transformers (ViTs) have achieved high classification accuracy, their large computational footprint makes deployment on resource constrained devices challenging. Existing compression approaches typically address pruning, quantization, and knowledge distillation in isolation, leaving the potential benefits and interactions of their combined application insufficiently explored. We propose a unified Vision Transformer compression framework that combines Hessian-Balanced Adaptive Block Pruning (H-BAC), guided by second-order sensitivity estimation, with quantization and attention-based knowledge distillation. To systematically identify the most effective configuration within each compression family, each technique is first evaluated independently through controlled ablation studies, after which the best-performing components are integrated into a sequential deployment pipeline tailored to real-world agricultural constraints. On a chilli 3-class village-split dataset with a genuine cross-village, cross-device out-of-distribution test split, the resulting compressed models match or exceed the 95.13% FP32 baseline's accuracy, alongside 74-98% model size reduction, and the fully integrated compression pipeline achieves a 54.5x size reduction (327.42 MB to 6.01 MB) at 95.13 +/- 2.32% accuracy across four tested configurations. A direct comparison further reveals that, on this dataset, a directly-trained student of the same final size, without pruning or distillation, reaches comparable accuracy of 94.87%, at the same 6.01 MB INT8 size, indicating where H-BAC and knowledge distillation are, and are not yet shown to be, worth their computational cost.
A runtime gate for an LLM tool agent is usually cast as a filter. In a ReAct loop a rejected proposal is followed by another at the same state, so the gate is a search operator over the proposal stream whose admission criterion shapes which trajectories are reachable. We study post-violation recovery admission, where progress must be admitted while the system is still in violation, and identify the scalar projection trap: an aggregate-score gate accepts a locally improving proposal and commits the trajectory to a plateau. SiLR instead shadow-executes each proposal and admits it under a product order over the branch-level violation state (overloaded-branch support and per-branch severity). We prove that no scalar surrogate is sound for this order, so the failure is representational, not a matter of threshold tuning. On mined Gym-ANM scenarios, SiLR recovers 21/21 multi-action episodes against 0/21 for terminal and 9/21 for the best scalar gate, significant across the full 24-scenario benchmark. The terminal-versus-structured dichotomy holds across three model families and in CityLearn. Because admission rests on deterministic simulation, the LLM lies outside the trust boundary: a magnitude-redistribution attack that defeats both scalar and support-only baselines is contained only by the full per-branch predicate. With two constraint families active, every tested scalar projection admits physically unsafe actions; support-only admits the largest fraction (63.2% of 42,410; product order 0). In the hardest dual-family traces, scalar gates recover only through that unsafe class. Reused as a GRPO process reward, it outperforms its count projection in every mined scenario and is the only tested reward whose ungated policy exceeds the untrained base (0.844 vs. 0.778). Scalar projection loses the violation geometry at both design points; only the full product order is structurally sufficient.
展开完整摘要收起摘要↓
A runtime gate for an LLM tool agent is usually cast as a filter. In a ReAct loop a rejected proposal is followed by another at the same state, so the gate is a search operator over the proposal stream whose admission criterion shapes which trajectories are reachable. We study post-violation recovery admission, where progress must be admitted while the system is still in violation, and identify the scalar projection trap: an aggregate-score gate accepts a locally improving proposal and commits the trajectory to a plateau. SiLR instead shadow-executes each proposal and admits it under a product order over the branch-level violation state (overloaded-branch support and per-branch severity). We prove that no scalar surrogate is sound for this order, so the failure is representational, not a matter of threshold tuning. On mined Gym-ANM scenarios, SiLR recovers 21/21 multi-action episodes against 0/21 for terminal and 9/21 for the best scalar gate, significant across the full 24-scenario benchmark. The terminal-versus-structured dichotomy holds across three model families and in CityLearn. Because admission rests on deterministic simulation, the LLM lies outside the trust boundary: a magnitude-redistribution attack that defeats both scalar and support-only baselines is contained only by the full per-branch predicate. With two constraint families active, every tested scalar projection admits physically unsafe actions; support-only admits the largest fraction (63.2% of 42,410; product order 0). In the hardest dual-family traces, scalar gates recover only through that unsafe class. Reused as a GRPO process reward, it outperforms its count projection in every mined scenario and is the only tested reward whose ungated policy exceeds the untrained base (0.844 vs. 0.778). Scalar projection loses the violation geometry at both design points; only the full product order is structurally sufficient.
作者Arno Libert, Derck W. E. Prinzhorn, Daan R. Henselmans
AI alignment requires AI systems to adhere to human norms, values, or intentions. Under value pluralism there is no correct target, but a shared prerequisite is that the system's behavior expresses a coherent policy: a mapping from situations to verdicts that is invariant while a situation's morally relevant features are preserved, and sensitive when they change. We introduce four structural conditions for such coherent policies: verdict stability, monotonicity, decisiveness, and Pareto viability. Together they measure a form of moral competence that is evaluable from behavior alone, without reference to a moral standard or expert baseline, forming a structural floor for alignment rather than a normative target. We demonstrate the methodology on three simulated deployments featuring LLM-based agents facing moral dilemmas. Evaluating nine frontier models under a factorial design of five paraphrases, five escalation levels, and three dominance conditions, we show no model expresses a coherent policy across the three deployments: surface-form perturbation alone produces verdict-rate shifts of up to $99$ percentage points at a single escalation level, and a model's success on one scenario does not predict its competence on another. This suggests LLM-based agents are not currently the kind of object to which alignment can meaningfully apply.
展开完整摘要收起摘要↓
AI alignment requires AI systems to adhere to human norms, values, or intentions. Under value pluralism there is no correct target, but a shared prerequisite is that the system's behavior expresses a coherent policy: a mapping from situations to verdicts that is invariant while a situation's morally relevant features are preserved, and sensitive when they change. We introduce four structural conditions for such coherent policies: verdict stability, monotonicity, decisiveness, and Pareto viability. Together they measure a form of moral competence that is evaluable from behavior alone, without reference to a moral standard or expert baseline, forming a structural floor for alignment rather than a normative target. We demonstrate the methodology on three simulated deployments featuring LLM-based agents facing moral dilemmas. Evaluating nine frontier models under a factorial design of five paraphrases, five escalation levels, and three dominance conditions, we show no model expresses a coherent policy across the three deployments: surface-form perturbation alone produces verdict-rate shifts of up to $99$ percentage points at a single escalation level, and a model's success on one scenario does not predict its competence on another. This suggests LLM-based agents are not currently the kind of object to which alignment can meaningfully apply.